Research map / Graph learning / Knowledge and temporal graphs
I Open at the Close: A Deep Reinforcement Learning Evaluation of Open Streets Initiatives
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Predicts NYC vehicle collisions with a recurrent GNN, then uses a deep Q-learning agent on a simulated Manhattan road network to choose which streets to open, finding Q-learning picks outperform NYC Open Streets picks and random streets.
Problem
Choosing which streets to open to pedestrians and cyclists (open streets) to reduce traffic congestion and vehicle collisions, which existing proposal processes do not simulate or evaluate for safety and congestion effects.
Approach
Frame the choice of open streets as a reinforcement learning problem: a recurrent GNN collision predictor supplies collision risk to the reward, traffic is simulated by rerouting taxi-derived trips around opened segments, and a deep Q-network values each segment's long-term reduction in collisions and congestion.
Key results
- The recurrent GNN achieved the best macro average recall of 0.76 ± 0.0040 on collision prediction, above XGBoost (0.74), LightGBM (0.72), and Graph WaveNet (0.64).
- Streets with the highest Q-values gave consistently positive reward, while NYC Open Streets program streets performed similarly to random streets with higher variance and lower average reward.
- Q-value-selected streets were more geographically spread across Manhattan, whereas program streets were concentrated in Downtown Manhattan.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Recurrent GNN instead of XGBoost for collision prediction | Collision prediction models · Collision prediction test set | macro average recall: 0.74$\pm$0.0001 -> 0.76$\pm$0.0040 | Helped |
| Use the full Recurrent GNN instead of the Lite variant (ablation) | Recurrent GNN · Collision prediction test set | macro average recall: 0.73$\pm$0.0043 -> 0.76$\pm$0.0040 | Helped |
| Select streets with the highest Q-values instead of streets selected by the NYC Open Streets program | Street-opening strategy · Simulated street openings | combined congestion and collision reward | Helped |
| Select streets with the highest Q-values instead of randomly selected streets | Street-opening strategy · Simulated street openings | combined congestion and collision reward | Helped |
| Select streets through the NYC Open Streets program instead of random selection | Street-opening strategy · Simulated street openings | combined congestion and collision reward | Hurt |
Surprising findings
- NYC Open Streets program streets had higher variance and a lower average reward than randomly selected streets.
- Most of the lowest-Q-value streets were East-West streets.
Limitations
- Taxi trips are rerouted locally after a street opens rather than drivers planning routes with knowledge of closures.
- Travel time is approximated by length times speed limit, ignoring real traffic.
- Only NYC/Manhattan is studied and objectives are limited to traffic and collisions.
Experimental setup
Datasets: NYC Open Data collision reports NYC TLC taxi trips 2013-2015 NYC road-bed infrastructure data NOAA weather data NYC Open Streets program street list
Benchmarks: Manhattan collision prediction test set (1,803,363 observations)
Metrics: macro average recall F1-score reward (collision risk and traffic reduction)
Baselines: Gaussian NB LightGBM XGBoost DSTGCN Graph WaveNet NYC Open Streets program selections random street selection
Builds on: Q-learning Graph WaveNet DSTGCN XGBoost LightGBM
Closest papers by meaning
- Learning to Follow Directions in Street View (AAAI 2020 · similarity 0.86)
- IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.84)
- Graph-Based Prediction and Planning Policy Network (GP3Net) for Scalable Self-Driving in Dynamic Environments Using Deep Reinforcement Learning (AAAI 2024 · similarity 0.84)
- π-Light: Programmatic Interpretable Reinforcement Learning for Resource-Limited Traffic Signal Control (AAAI 2024 · similarity 0.83)
- URB - Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles (NeurIPS 2025 · similarity 0.83)
- DiffLight: A Partial Rewards Conditioned Diffusion Model for Traffic Signal Control with Missing Data (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.